Date of Award
Doctor of Philosophy (PhD)
As digitized clinical and health data become ubiquitous, machine learning techniques have shown promise in predicting various clinical outcomes. In this thesis research, we exploit three types of data including (1) data collected through wearables outside hospitals, (2) electronic health records (EHR) data of inpatient in general hospital wards, (3) intraoperative data collected during surgery. This thesis work investigates machine learning approaches for the diverse clinical and health data with distinctive characteristics and challenges in the context of real-world clinical applications. Specifically, this thesis makes the following contributions to the state of the art of clinical machine learning.
Extracting informative features from wearable data for clinical monitoring. Wearable devices (e.g., smart wristbands and watches) provide unprecedented capabilities to continuously monitor patients unobtrusively outside hospitals. However, we face the challenge to extract informative features from fine-grained time-series data (e.g., steps, heart rate, sleep stages) collected by wearables. We developed feature engineering approaches specifically tailored for wearable data for the purpose of predicting clinical outcomes. We developed predictive models and evaluated our feature engineering and modeling approaches in the context of two clinical studies using Fitbit wristbands: (1) predicting clinical deterioration and readmissions of congestive heart failure patients, and (2) predicting postoperative complications and readmissions of patients subject to pancreatectomy. Multi-horizon Alerts for Clinical Deterioration. Clinical early warning systems (EWS) aim to predict clinical deterioration of hospitalized patients in general hospital wards. To leverage the alerts for effective clinical intervention and planning, we propose to generate alerts with various predictive time horizons indicating different levels of urgency. We proposed DeepAlerts, a deep multi-task model to predict deterioration at different future horizons. The novelties of DeepAlerts lie in its application of prior knowledge regularization to the deep multi-task learning and a task-specific loss balancing for multi-task learning. We applied DeepAlerts to an EWS model based on the EHR data from 20,700 hospitalizations of adult oncology patients. Integrating Static and Time-series Data for Clinical Early Warning. The static and time-series data in EHR often make complementary contributions to clinical predictions. To effectively integrate static and time-series data, we developed CrossNet, a deep predictive model that employed multi-modal fusion to integrate static and time-series variables in deep recurrent models and introduced cross-modal imputation to improve the predictive outcomes. CrossNet was validated on the EHR data from hospitalized adult oncology patients in the context of EWS and MIMIC-III ICU dataset for generalizability.
Self-explaining Hierarchical Model for Intraoperative Time Series. Postoperative complications are potentially preventable via early predictions based on intraoperative time series. To overcome the challenges posed by long time series with large gaps in surgeries, we proposed a Self-Explaining Hierarchical Model (SEHM) combining the strength of locality-based attention and recurrent models. In addition, SEHM ensures end-to-end interpretability pinpointing data points with potential clinical significance. Experiments on a large dataset of 111,888 surgeries with multiple outcomes show that our model can achieve strong predictive performance and offer robust interpretations for predicted outcomes based on intraoperative time series.
Available for download on Thursday, June 01, 2023